摘要
目的利用机器学习算法辅助建立急性心力衰竭(AHF)患者中、短期生存预测模型,分析其预后影响因素并进行验证。方法回顾性分析北京医院急诊科2021年7月至2025年2月收治的AHF患者的临床资料,包括人口统计学信息、基础疾病、生命体征、合并症、7天以内的实验室检查指标、诊断和治疗等,根据随访90天的临床结局分为存活组和死亡组。采用L1正则化的Logistic回归分析、随机森林法和极限梯度提升法纳入临床资料及其一阶数据,分别以生存结局建立预测模型,评估模型预测效能;进行内部验证,并筛选影响预后的主要临床因素,同时分析影响不同时点生存状态的因素差异,随后对模型中排名靠前的参数进行定量分析及生存分析以量化其预测效能。结果基于就诊后90天的生存状况将患者分为存活组102例,死亡组66例,存活组和死亡组的单变量差异有统计学意义。在90天预后模型中随机森林法预测效果最佳,曲线下面积(AUC)为0.782,准确率为0.667,敏感度为0.838,特异度为0.600,模型重要性前五位的因素依次为第7天的D-二聚体、第3天的舒张压、第7天和第5天的肌钙蛋白I和第7天的N-末端脑钠肽前体(NT-proBNP)。进一步探索发现90天的其他模型以及30、60、90天的预测模型中重要性排名前列的因素有较高的重合性,第7天的尿素氮和D-二聚体在多个模型中排名均位于前两位。然后采用Logistic回归分析量化第7天的尿素氮和D-二聚体对90天生存的预测价值,发现两者均有独立预测价值,且联合预测价值更大。生存分析发现两参数的水平对生存率有显著影响。结论本研究通过机器学习算法,基于常见临床因素建立了AHF患者90天生存预测模型,并进一步量化和验证关键指标对生存率的预测效能。模型显示,AHF患者入院后第7天的尿素氮和D-二聚体对90天生存预后具有较高的预测价值,在更短期的预后(30天和60天)预测中有类似趋势。
Objective To establish a medium-and short-term survival prediction model for the patients with acute heart failure(AHF)assisted by machine learning algorithms and to analyze and verify the prognostic factors.Methods The clinical data of AHF patients admitted to the Emergency Department of Beijing Hospital from July 2021 to February 2025 were retrospectively analyzed,including demographic information,underlying diseases,vital signs,comorbidities,laboratory test indicators,diagnosis and treatment within 7 days,etc.According to the clinical outcome of 90-day follow-up,they were divided into survival group and death group.L1 regularized Logistic regression analysis,random forest method and extreme gradient boosting method were used in the training set,and clinical data and its first-order data were included.Prediction models were established based on survival outcomes.Predictive efficiency of the model was evaluated,internal verification was conducted in the test set,and the main clinical factors affecting prognosis were screened,and the differences in the factors affecting the survival status of different times were analyzed,then quantitative analysis and survival analysis were performed on the top-ranked parameters in the model to quantify their predictive efficiency.Results Based on the survival status 90 days after the treatment,the patients were divided into the survival group with 102 cases and the death group with 66 cases.There were significant differences in the univariate between the survival group and the death group.The random forest method showed the best predictive results in the 90-day prognosis model,with the area under curve(AUC)of 0.782,the accuracy was 0.667,the sensitivity was 0.838,and the specificity was 0.600.The top five factors in the importance of the model were D-dimer on 7th day,diastolic blood pressure on 3rd day,troponin I on 7th and 5th day,and N-terminal probrain natriuretic peptide(NT-proBNP)on 7th day.Further exploration found that the factors ranked the top importance in other models of 90 days and the prediction models of 30,60 and 90 days had high overlap.Urea nitrogen and D-dimer on 7th day were ranked the top two in multiple models.Then,the predictive value of urea nitrogen and D-dimer on 7th day for 90-day survival was quantified by Logistic regression analysis,and it was found that both had independent predictive value and the combined predictive value was greater.Survival analysis found that the level of two parameters had a significant impact on survival rate.Conclusions This study establishs a 90-day survival prediction model for the patients with AHF based on common clinical factors through machine learning algorithms,and further quantifies and verifies the good predictive efficacy of key indicators for survival.The model shows that urea nitrogen and D-dimer on 7th day after admission for AHF patients had high predictive value in their 90-day survival prognosis,with similar trends in the shorter-term prognosis(30-day and 60-day)predictions.
作者
郑亮亮
王俊杰
王凡
全锦花
陈曦
温伟
Zheng Liangliang;Wang Junjie;Wang Fan;Quan Jinhua;Chen Xi;Wen Wei(Emergency Department of Beijing Hospital,National Center of Gerontology,Institute of Geriatric Medicine,Chinese Academy of Medical Sciences,Beijing 100730,China)
出处
《中国急救医学》
2025年第7期595-602,共8页
Chinese Journal of Critical Care Medicine
基金
社会团体基金<不同疗程重组人脑利钠肽治疗急性心力衰竭的疗效差异及安全性比较>(BJ-2022-203)。
关键词
急性心力衰竭
机器学习
生存分析
预测模型
Acute heart failure
Machine learning
Survival analysis
Predictive model
